Entity Recognition, Linking, and Relation Extraction for Cardiovascular and Cerebrovascular Diseases and Therapeutic Drugs in EHR CLinical Notes

LLaMIC is a modular pipeline designed for clinical entity recognition and standardization, focusing on identifying cardiovascular and cerebrovascular disease (CCD) mentions, their therapeutic drugs (CCDt), and binary relations in unstructured electronic health records (EHR) clinical notes.

LLaMIC combines instruction-tuned large language models (LLMs) with rule-based lexicons to achieve:

• Named Entity Recognition (NER): Recognition of CCD mentions within ICD-10 codes I20–I25 (ischemic heart diseases) and I60–I64 (cerebrovascular diseases), as well as CCDt mentions.
• Named Entity Linking (NEL): Linking recognized entities to standardized ICD-10/MeSH concepts.
• Relation Extraction (RE): Extracting binary relations between disease and drug entities.

The first two tasks are performed by LLaMIC-Entities, while the last is handled by LLaMIC-RE.

Clone the repository and install required dependencies in your environment:

pip install -r requirements.txt

Set the model and lexicon paths in call_llamic.sh: Select the entity type (disease or drug) and specify the LLaMA model to use for the NER, NEL, and review phases. You can choose the Meta-Llama-3.1-8B-Instruct model (llama3.1) or provide the path to your own fine-tuned checkpoints. Note that access to Meta-Llama-3.1-8B-Instruct must first be requested from Hugging Face, as it is restricted (https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct).

Update the domain lexicon (lexicon.csv) with the target ICD-10 codes for diseases or MeSH codes and descriptions for drugs. Adjust the n_iterations parameter for NER if needed.

--entity_type disease (disease or drug)
--model_name_or_path_ner llama3.1 (default)
--model_name_or_path_nel llama3.1 (default)
--model_name_or_path_review llama3.1 (default)

--lexicon_path $datadir/lexicon.csv
--n_iterations 1 (default)

Run the LLaMIC on HPC with SLURM:

sbatch call_llamic.sh

The fine-tuned LAMIC model was evaluated on cardiovascular and cerebrovascular diseases and ICD-10 entity linking tasks. In MIMIC-IV notes dataset, it achieved precision scores of 0.83 (strict mode) and 0.89 (lenient mode) for disease identification. For ICD-10 linking, the model reached a precision of approximately 0.82.

Clone the repository and install required dependencies in your environment:

pip install -r requirements.txt

Set the model paths in call_llamic.sh: You can choose the Meta-Llama-3.1-8B-Instruct model (llama3.1) or provide the path to your own fine-tuned checkpoints. Note that access to Meta-Llama-3.1-8B-Instruct must first be requested from Hugging Face, as it is restricted (https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct).

--model_name_or_path_pg llama3.1 (default)
--model_name_or_path_rc llama3.1 (default)

Run the LLaMIC on HPC with SLURM:

sbatch call_llamic.sh

The fine-tuned LLaMIC model was evaluated on the MIMIC-IV notes dataset. In the PG phase (Pair Generation), it achieved a precision score of 0.67.

https://physionet.org/content/mimic-iv-note/2.2/